The first approach consisted of "dissection" of the complex AF trait by grouping the offspring in accordance with their 660 gene transcriptome profiles. Schadt et al.  were the first to use this strategy and they improved the significance of a fatness QTL previously detected on a chromosome and even detected a new QTL on another. We found similar results in the present study. First, we found an increase in the LRT of GGA5 AF QTL by using subgroups 1 and 4. Second, despite the small size of our experimental design (45 animals), we detected using subgroups 2 and 3 an AF QTL at 102 cM on GGA5, as previously detected using an experimental design with 1300 birds . These results show the power of the approach. In view of the polygenic influence on the complex traits, the variations in abdominal fatness are probably due to variations in several biological pathways, impacted by multiple mutations acting separately or in interaction. Transcriptome data offer the possibility of dissecting such a complex trait in more elementary phenotypes (gene expressions correlated with the trait) and therefore make it possible to separate the population into genetically homogenous subgroups using transcriptome profiling. Some combinations of subgroups possibly reflect the effects of a precise mutation whereas others reflect the signature of other mutations. In this study, we showed the over representation in some subgroups of birds for which the AF value was in disagreement with the paternal Q/q haplotype of the two GGA5 AF QTL, thus increasing the QTL detection power when they were removed. These results indicate that the QTL each significantly affect only a subset of the population analyzed. This heterogeneity observed between offspring clearly demonstrates the complexity underlying traits such as fatness. In summary, the 660 genes correlated with the AF allowed classification of offspring in a relevant way to dissect AF QTL on the GGA5 chromosome, despite the small number of animals analyzed.
To the best of our knowledge, no study using this approach has been published since the first publication in 2003 . Our results obtained with a livestock design and those of Schadt et al. obtained with mice indicate that the identification of subgroups in a population on the basis of transcriptome profiles would be an effective way to improve the power of QTL detection by linkage analysis.
The second strategy aimed at improving characterization of GGA5 AF QTL by eQTL mapping. Such a strategy is widely used in the context of QTL detection of complex traits using transcriptome data [10, 11, 19, 23, 30–37]. The principle is to identify genes correlated with the complex trait that have an eQTL co-localizing with the QTL of interest. Most of these authors then focused on genes with a function related to the complex trait and having a cis-eQTL, allowing them to hypothesize that the mutation responsible for the complex trait is in the cis-eQTL gene [10, 30, 32, 34–37]. In our study, no cis-eQTL was detected among the 46 genes correlated with the AF trait and having an eQTL in the GGA5 AF QTL region. Correlation pairs between these genes strongly suggested that these 46 trans-eQTL gene expressions were probably controlled by different mutations in the location confidence interval of the GGA5 AF QTL, as previously commented by Georges  and Schadt et al. . Because the CI of the QTL was large (31 cM), we therefore selected 5 genes that effectively corrected the AF QTL (P > 0.1), and therefore likely to be controlled by the mutation sought or by a mutation close to it. Functional analysis of these 5 genes was still limited by the partial functional annotation of genes. However, we identified one gene related to lipid metabolism that could be affected by the mutation sought. This gene encodes HGMCS1, known to be involved in cholesterol metabolism. We have recently shown that its regulation in response to fasting is different in chickens compared to mammals . Further experiments will be necessary to clarify its role in fatty acid metabolism and its regulation in chickens in order to target a potential regulatory gene in the distal GGA5 AF QTL.
Moreover, the whole 5 gene-set considered as a signature of the mutation underlying the QTL of interest or a mutation close to it may be useful to refine this QTL using a multivariate model that takes advantage of the correlation between these 5 expression traits and AF. Multivariate analysis combining the CV5 variable and the AF trait led to a significant increase in maximum LRT compared to the AF trait. This result supported the hypotheses of the existence of QTL affecting both AF and CV5 at the same position or the existence of different close mutations in linkage disequilibrium. We were unable to reach a conclusion with such a small number of animals analyzed. However, this result makes it possible to reduce the location confidence interval of GGA5 AF QTL from 156-187 cM (31 cM) to 166-184 cM (18 cM).
Finally, an original approach to refine a QTL region was proposed in this study. We used the same 5 gene-set to find the best gene expression combination discriminating the paternal Q from q haplotypes (corresponding to the whole confidence interval of the GGA5 AF QTL) and used it to predict the Q versus q mutation received by the recombinant animals. Genotyping these birds with additional markers drastically reduced the region to 166-173 cM (7 cM). Contrary to conventional approaches used to refine a QTL, such a strategy avoided generating new offspring to test the QTL genotype of the recombinant birds and saved on high levels of genotyping, thus gaining time and saving money. However, it is important to remember that it was based on the relevance of the gene-set considered as the signature of the QTL mutation or mutations close to it.
The gains in power and precision of QTL detection offered by approaches 1 and 2 were probably limited by the low density of markers and size of the experimental design used in this study (45 birds). Nevertheless, these approaches allowed substantial reduction of the GGA5 AF QTL region (20% up to 50%). Approach 3 was more effective (80% reduction), depending on the recombination breakpoints in the recombinant birds. This third approach allowed us to refine the GGA5 AF QTL from 156-187 cM (31 cM) to a most probable location confidence interval of 166-173 cM (7 cM) (Table 1). It can be seen that this reduction of GGA5 AF QTL region was consistent with the other more limited reductions obtained by the other two approaches. Unfortunately, a gene by gene bibliography analysis did not allow us to propose a good functional and positional gene candidate as regulator of HMGCS1 and the AF trait.